Toward automated detection and diagnosis of mammographic microcalcifications

Mammographic diagnosis is the most effective technique to detect breast cancer in its infancy when it is most responsive to treatment. An early and a significant indicator of breast cancer is the presence of clustered microcalcifications (MCs). Mammographic MCs greatly vary in their appearance and shape, and become indistinguishable when surrounded by dense breast tissue. This makes radiologist's interpretation of mammograms a tedious and an error prone task.;Although computer aided diagnosis (CAD) methods are being developed to aid radiologist in detecting and analyzing the malignancy of MCs, existing systems have not achieved a satisfactory performance. The specificity of existing methods is low compared to a radiologist's interpretation. Therefore, there is a need for exploring new detection methods and developing automated, robust feature extraction and selection techniques that support the diagnosis process.;To address these needs, a detection framework that employs a pattern-synthesizing process along with statistical and spectral characterization of mammograms is proposed. A trained statistical Bayesian classifier using synthetic MCs will be used to classify anonymous input patterns into a background or microcalcification classes. Morphological image processing is also proposed in this dissertation to segment and characterize the shape and the distribution of MCs. Automated nested subsets feature selection method and heuristic search method are investigated via a full model selection using PSO-SVM framework. Furthermore, a new approach to extract texture features of MCs using a multiscale Hessian image analysis is developed and tested.;The detection and diagnosis schemes developed in this dissertation are tested using mammograms from the Mammographic Image Analysis Society (MIAS) database and compared to other existing methods. The results indicate that the performance of the detection scheme is adequate while the performance of the shape-based diagnosis of MCs scheme is superior and very promising.